Query Fan-Out: What 9,238 ChatGPT Searches Show About AI Search

Disclosure: This article uses data collected with querying.ai, our own product, and shows its API. Read more about Switchbreak.

Query fan-out is the step where an AI search engine turns one question into several narrower web searches, runs them, and writes a single answer from what comes back. Google says AI Mode and AI Overviews may work this way, and OpenAI says ChatGPT search typically rewrites a question into one or more targeted queries for its search providers. If you want your pages to appear in AI answers, those searches matter more than the question itself, because they are what the engine actually looks up.

We logged the searches ChatGPT ran while answering the same 150 questions every day for 28 days: 9,238 fan-out queries in all. This article explains query fan-out, shows what those queries looked like, and covers how to see them for your own topic and what to change once you can.

What Google and OpenAI say about query fan-out

Google's May 2025 I/O post on AI Mode said: "Under the hood, AI Mode uses our query fan-out technique, breaking down your question into subtopics and issuing a multitude of queries simultaneously on your behalf." Deep Search, for more thorough answers, "uses the same query fan-out technique but taken to the next level" and "can issue hundreds of searches."

Google's guidance for site owners applies the idea to both features: AI Overviews and AI Mode "may use a 'query fan-out' technique — issuing multiple related searches across subtopics and data sources — to develop a response." Google says this helps them show "a wider and more diverse set of helpful links."

For ChatGPT, OpenAI's help article on ChatGPT search says it "typically rewrites your query into one or more targeted queries" and sends them to its search providers. In OpenAI's example, a question about drugs that target CCR8 for cancer might first become the search "CCR8 immunotherapy drug development 2025", followed by more specific searches. The same page says that, if memory is enabled, ChatGPT may use saved memories when it rewrites a search query, and that it may use an approximate location based on your IP address for local results. Two people asking the same question can trigger different searches.

Fan-out does not come with a new rulebook. Google's guidance says, "There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary." A page still has to be indexed and "eligible to be shown in Google Search with a snippet" to appear as a supporting link.

One question, two different fan-outs

On October 16, 2026, we asked ChatGPT "why do ai search engines rewrite my prompts?" It ran three searches:

• "AI search engines query rewriting prompt rewriting query fan out official documentation"
• "Google AI search query fan-out official documentation"
• "Perplexity search query rewriting queries documentation"

On October 15, the same question produced a different set:

• "AI search engines rewrite user queries query rewriting query expansion official documentation"
• "Google AI Mode query fan-out searches query rewrite official"
• "Bing search query reformulation AI search official documentation"

Three things show up in both runs. Besides one general search, each search focused on a single search engine. It added words the person never typed, such as "official documentation." And it changed its searches from one day to the next: the third search went to Perplexity on one day and to Bing on the other.

How we measured

• Questions. 150 English questions about AI search tracking tools and APIs, asked from the United States. None of them mentions a year.
• Engines and dates. ChatGPT, Gemini and Google AI Overviews, once a day from September 21 to October 18, 2026, for 12,595 answers in all.
• Fan-out data. Only ChatGPT returned the searches it ran, so every query figure below comes from its 4,195 answers. Gemini and AI Overviews returned answers and links but no list of searches.
• Names. The questions are about our own market, so many searches name companies in it. We call the three vendors named most often across the three engines Vendor A, B and C.
• A change mid-period. From October 9, the daily share of ChatGPT answers with recorded searches rose from about 93% to about 99%, which coincides with a change in how we collect. The figures cover all 28 days.

What 9,238 fan-out queries looked like

ChatGPT searched almost every time, a few searches at a time

95.0% of ChatGPT answers (3,986 of 4,195) came with at least one recorded web search. Answers that searched ran 2.32 searches on average. The median was 2, nine in ten ran 4 or fewer, and the most in one answer was 12.

Longer questions fanned out more

The 127 questions longer than six words led to a search in 96.4% of answers and averaged 2.33 searches per answer. The 23 questions of six words or fewer led to a search in 87.4% of answers and averaged 1.52.

The searches stayed about ten words long either way: 10.4 words for questions averaging 23 words, and 10.1 words for questions averaging 5. ChatGPT condensed long questions and expanded short ones. Only 28 of the 9,238 searches (0.3%) repeated the question word for word.

What ChatGPT added to the question

The searches read like research notes. 31.9% of answers that searched included at least one search naming a vendor, and 37.3% included a year, though none of the questions did. Words for ranking or comparing options were rare: "best" or "top" appeared in 4.0% of searches, and comparison words in under 1%.

Every run searched differently

Over the 28 days, a single question produced a median of 54 different searches, with a range of 11 to 110. A question's most common search showed up in a median of 10.7% of its answers. The 9,238 searches contained 7,573 distinct strings. A single answer shows you one draw from a wide spread, so read fan-out across many runs.

Vendors in the searches matched vendors in the answers

When a search named a vendor, the answer usually named that vendor too. The table shows how often the answer named each vendor, among the 3,986 ChatGPT answers that searched:

This is a correlation, and it may run either way: ChatGPT may search for a vendor because it already plans to mention it. Still, the searches suggest which names ChatGPT is considering for the answer.

Site-restricted searches show what ChatGPT thinks your domain is

In 423 answers (10.6% of those that searched), ChatGPT limited at least one search to a single website with the site: operator. Those answers ran 4.57 searches on average, 3.26 of them restricted to a site.

For two vendors whose brand name matches their domain, these searches went to the right domain, 276 and 274 times. For a vendor whose main domain differs from its brand name, they spread across about 20 host names. The most common was a misspelling of the brand, used 113 times. The domain that answers cited most often for that vendor was searched only 30 times.

This data can't say why ChatGPT guessed wrong. It does point to something you control: if your domain differs from your brand name, or your name is easy to misspell, state the connection plainly and consistently wherever you are described.

Keyword tools can't see fan-out queries

We checked ten of the most repeated searches, each run 17 or 18 times over the 28 days, against Google Ads data for the United States. All ten showed 0 average monthly searches. Fan-out queries are long, specific and generated on the fly, so keyword volume won't tell you which ones matter. You have to collect them from the answers.

How to see the fan-out queries for your topic

In querying.ai, these engines return the searches they ran:

• ChatGPT (CHATGPT): searchQueries lists the searches. Each entry in sources also has its own searchQueries, the searches whose results included that page, and cited, which says whether the answer cites the page inline. Both arrive by default.
• Perplexity (PERPLEXITY): set include.searchQueries to true to get searchQueries.
• Bing Copilot (BING_COPILOT): searchQueries is always present, and empty when Copilot produced none. Meta AI, Qwen and Doubao return searchQueries too.
• Gemini, Google AI Overviews and AI Mode return answers and links but no list of searches.

Submit a question to ChatGPT:

curl -X POST "https://api.querying.ai/v1/async/task" \
  -H "Authorization: Bearer $QUERYING_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "taskType": "CHATGPT",
    "payload": {
      "prompt": "which crm is easiest for a small team",
      "country": "US"
    }
  }'

The reply includes the task's id. Poll GET /v1/async/task/:id until its status is COMPLETED, or add a webhook.url to receive the result as soon as it is ready. A finished task can be fetched for 24 hours, so save each result when it arrives.

This function takes the task's response and lists each search with the number of pages in the source list that came from it and how many of those the answer cited:

function fanOutReport(response) {
  const rows = new Map();
  for (const query of response.searchQueries ?? []) {
    rows.set(query, { pages: 0, cited: 0 });
  }
  for (const source of response.sources ?? []) {
    for (const query of source.searchQueries ?? []) {
      const row = rows.get(query) ?? { pages: 0, cited: 0 };
      row.pages += 1;
      if (source.cited) row.cited += 1;
      rows.set(query, row);
    }
  }
  return [...rows].map(([query, row]) => ({ query, ...row }));
}

Run your questions on a schedule and keep every report. Three counts are worth tracking: how often each search, or each recurring word, appears for a question; which of your pages appear under which searches; and how often those pages are cited. A page that keeps turning up under a search but is rarely cited is being found and passed over.

To check whether a fan-out query has any search demand, send it to KEYWORD_VOLUME. It takes up to 100 keywords per task, each up to 80 characters, for 3 credits, and returns the average monthly searches over the last 12 months. Expect many zeros.

What to do with fan-out queries

1. Cover the subtopics the searches reveal. If the searches for your topic keep splitting into pricing, documentation, integrations and limits, each one needs a clear answer on a page that says what it covers.
2. Keep an official page current and dated. One search in eight included the word "official," and more than a third of answers that searched included a year. A dated, official page gives those searches something to find.
3. Make your name and domain easy to connect. Use one spelling of your brand, link it to your domain on your own pages and profiles, and keep the two consistent in your structured data.
4. Stay indexable and eligible for snippets. For Google, that is the stated requirement for appearing as a supporting link, and fan-out doesn't change it.
5. Don't publish a page per query. With 7,573 distinct search strings in four weeks for one topic, it isn't possible anyway, and Google's spam policies describe the risk: "Scaled content abuse is when many pages are generated for the primary purpose of manipulating search rankings and not helping users." One of their examples is using generative AI tools "to generate many pages without adding value for users."
6. Measure across many runs. Fan-out changes from one run to the next, so judge your coverage by rates over weeks.

Limits of this data

• It covers one market, AI search tracking tools, in English from the United States. Other topics may fan out differently.
• Only ChatGPT exposed its searches. Gemini and AI Overviews returned no list of searches, so this data says nothing about how Google fans out.
• The link between vendors in searches and vendors in answers is a correlation.
• The collection change on October 9 raised the recorded share of ChatGPT answers with searches from about 93% to about 99% a day, and the totals mix both periods.
• ChatGPT's searches and results can vary with saved memories and location, so other people may see different fan-out for the same question.

Cited or Just Listed shows how to separate the pages an answer cites from the ones it only lists, and AI Share of Voice uses the same answers to compare ways of scoring brand mentions.